Identifying the top predictors of student well-being across cultures using machine learning and conventional statistics

Alongside academic learning, there is increasing recognition that educational systems must also cater to students’ well-being. This study examines the key factors that predict adolescent students’ subjective well-being, indexed by life satisfaction, positive affect, and negative affect. Data from 522,836 secondary school students from 71 countries/regions across eight different cultural contexts were analyzed. Underpinned by Bronfenbrenner’s bioecological theory, both machine learning (i.e., light gradient-boosting machine) and conventional statistics (i.e., hierarchical linear modeling) were used to examine the roles of person, process, and context factors. Among the multiple predictors examined, school belonging and sense of meaning emerged as the common predictors of the various well-being dimensions. Different well-being dimensions also had distinct predictors. Life satisfaction was best predicted by a sense of meaning, school belonging, parental support, fear of failure, and GDP per capita. Positive affect was most strongly predicted by resilience, sense of meaning, school belonging, parental support, and GDP per capita. Negative affect was most strongly predicted by fear of failure, gender, being bullied, school belonging, and sense of meaning. There was a remarkable level of cross-cultural similarity in terms of the top predictors of well-being across the globe. Theoretical and practical implications are discussed.


Bronfenbrenner's bioecological theory
To address the role of multiple factors in well-being, Bronfenbrenner's bioecological theory was used.It is one of the most prominent and comprehensive frameworks that can be used to understand human functioning 23 .It has also been used in prior research to examine well-being 24 .The bioecological theory focuses on the role of four key factors that shape human development: proximal processes, person, context, and time 20,[25][26][27] .We elucidate these factors below.

Proximal processes
Proximal processes involve reciprocal interactions between individuals and their social partners (i.e., people, symbols, tasks, and objects).They are the primary mechanisms of interactions between humans and the environment.In this study, proximal processes pertain to how students engage with their learning materials and academic activities, as demonstrated by their usage of meta-cognitive strategies (i.e., summarizing and understanding, memorizing, and assessing credibility) 7,28 .Meta-cognitive strategies have been found to be positively associated with subjective well-being in previous studies 29,30 .

Person factors
Person factors include innate characteristics such as demographic characteristics (e.g., gender), personality traits, motivation, and attitudes.Gender has been linked to well-being, with girls experiencing higher negative affect than boys 31,32 .Goals and aspirations are also key factors in understanding students' well-being 33,34 .Individuals who made better progress toward their goals or those who are able to realize their aspirations have higher levels of well-being than others 34 .
Other psychological factors such as self-efficacy when facing adversity, fear of failure, competitiveness, as well as self-concept of task difficulty and competence might also be closely associated with well-being.Together, these concepts capture students' beliefs in their capability to cope with adversity and other challenging situations [35][36][37] .Self-efficacy and self-concept have positive associations with well-being 36,38 , while fear of failure undermines well-being 39,40 .

Context factors
Context pertains to the physical and social environment and can be further divided into microsystem, mesosystem, exosystem, and macrosystem levels.The microsystem is the immediate setting in which an individual lives and can include the family and school contexts.The mesosystem refers to the interconnections between different components of the microsystem (e.g., the connection between a child's family and the school).The exosystem includes contexts that indirectly impact the individual's development even if he/she is not directly participating in it.For example, a parent's workplace can impact a child's socio-emotional adjustment.The broadest is

Time
Time incorporates multiple time scales of development and captures individuals' trajectories.Time was not included in this study due to the cross-sectional nature of the PISA dataset.It is important to note, however, that PISA focuses on 15-year-old adolescent students, and the findings of this study are situated within this developmental stage.
Despite some cross-cultural differences, the expectations for adolescent students across the globe share certain similarities 55 .Adolescent students are expected to do well in school and prepare for either going into higher education or joining the workforce after secondary education.Adolescence is also a critical period for social and emotional development, and students are expected to develop healthy relationships and self-awareness, while navigating the biological and social changes associated with puberty.These societal expectations could shape adolescents' well-being.

Cultural similarities and differences
Well-being varies across cultures 56 .However, much of the current research on well-being has mostly relied on Western samples.Culture involves a rich complexity of "meanings, beliefs, practices, symbols, norms, and values prevalent among people in a society" 57 .Schwartz proposed the Cultural Values Theory to explore how different cultures vary in terms of their value orientations 57 .He proposed that different societies across the world can be categorized into eight distinct cultures based on how they prioritize cultural values.
The first dimension of cultural value contrasts autonomy (emphasis on creativity, curiosity, self-expression, pleasure, and enjoyment) with embeddedness (emphasis on social hierarchy, authority, and respect for tradition).The second dimension contrasts hierarchy (emphasis on social hierarchy, authority, and tradition) with egalitarianism (emphasis on equality, fairness, and justice), and the third dimension contrasts mastery (emphasis on achievement, success, and competence) with harmony (emphasis on social relationships, mutual respect, and consensus).
Based on how countries prioritize different cultural values, they can be classified into eight cultural groups: Africa and the Middle East, Confucian, East-Central Europe, East Europe, English Speaking, Latin America, Southeast Asia, and West Europe 58 .For example, Confucian Asia (e.g., China) is high in embeddedness, hierarchy, and mastery.Countries in Africa and the Middle East (e.g., Nigeria) score higher in embeddedness and have lower scores in mastery and autonomy.

Explanation and prediction paradigms
In analyzing the data for this study, we use both the explanation and prediction paradigms.Explanation focuses on describing the causal relationships among variables by drawing on specific theoretical models.Conventional statistics is typically rooted in the explanation paradigm.It is usually grounded in a parsimonious theoretical model and can be used to explore the relationship between the independent and dependent variables 40,42 .Conventional statistics has the advantage of generating interpretable parameter estimates.For example, one can use conventional statistics (e.g., linear regression) to estimate the direction and magnitude strength of the association between a predictor (X) and an outcome variable (Y).The researcher can input data for the independent and dependent variables into the regression model and generate a parameter estimate that captures the direction and magnitude of the association between X and Y.
Machine learning, on the other hand, is rooted in the prediction paradigm.It does not generate parameter estimates and is a 'black box' .Instead, machine learning focuses on identifying the most powerful predictors of the outcome variables.For example, a researcher using machine learning can input 100 predictors into the model and let the machine identify which among the variables best predict the outcome.By leveraging advanced algorithms, machine learning enables researchers to delve into large-scale datasets and uncover patterns in the data that would otherwise not have been possible with conventional statistics 59 .
Compared to conventional statistical methods, machine learning methods provide flexibility for modeling a large number of predictors and complex associations (i.e., nonlinearity and interaction) between predictors and outcomes 59 .Unlike conventional statistics, it can handle highly correlated predictors.In addition, machine learning involves splitting the data into a training set and a validation set.This maximizes the generalizability of findings to new data, optimizes predictive accuracy, and reduces problems of overfitting 60 .However, machine learning results are not readily interpretable, as they do not generate interpretable parameter estimates such as beta coefficients.Hence, in this study, we aimed to use both machine learning and conventional statistical analyses.

The present study
In the current study, we aimed to (1) identify the most important predictors of students' subjective well-being using machine learning approaches (prediction) and (2) explore how these predictors contributed to explaining variance in students' subjective well-being using conventional statistics (explanation).Hence, we drew on both the prediction paradigm of machine learning and the explanation paradigm of conventional statistics and leveraged the strengths of both approaches.
We also examined how the patterns of relationship between the predictors and subjective well-being outcomes were similar or different across cultural contexts (i.e., Africa and the Middle East, Confucian, East-Central Europe, East Europe, English Speaking, Latin America, Southeast Asia, and West Europe).The conceptual framework for the present study is shown in Fig. 1.   1 shows the countries and sample size of each cultural group.Ethical approval was not required for this study as we used secondary analyses of existing data that is publicly available and de-identified.

Subjective well-being
Subjective well-being was the key dependent variable.It was operationalized in terms of students' life satisfaction, positive affect, and negative affect.Students were asked about their overall life satisfaction with one item (i.e., "Overall, how satisfied are you with your life as a whole these days").This item was rated from 0 to 10, with higher numbers representing a higher level of life satisfaction.
Positive and negative affect were operationalized as how they generally feel in their lives, using five positive adjectives (e.g., joyful) and four negative adjectives (e.g., afraid), each of which was rated on a 4-point Likert scale (1 = Never to 4 = Always).The internal consistencies for positive affect (Cronbach's α = 0.79) and negative affect (Cronbach's α = 0.74) were acceptable.

Predictors
Based on Bronfenbrenner's bioecological theory 23 , 37 variables were selected from the PISA dataset as predictor variables (see Table 2 for detailed descriptions of all variables).These predictors were based on the PISA Assessment and Analytical Framework created by the OECD 7 (see https:// www.oecd.org/ educa tion/ pisa-2018-asses sment-and-analy tical-frame work-b25ef ab8-en.htm).PISA encompasses many items/variables related to students' , parents' , and schools' characteristics.Using these items/variables, OECD calculated derived variables based on item response theory (IRT) scaling.Given that the focus of PISA 2018 was on student well-being, many of the variables in the database were specifically selected by the OECD because of their theoretical linkages to wellbeing in the existing literature.
Two additional country factors (i.e., Gini and GDP per capita) were used from the World Bank website (https:// www.world bank.org/ en/ home).The Cronbach's alpha internal reliability values of these independent variables ranged from 0.64 to 0.91.

Analysis
In the preliminary analysis, we excluded 9 countries that had high rates of missing data, ranging from 18.9% to 44.0%.The excluded countries were Norway, Belgium, North Macedonia, Mexico, Australia, New Zealand, Canada, Singapore, and Israel.Next, we clustered the remaining 71 countries/regions into Schwartz's eight cultural groups 58 .Missing data were imputed using the missForest package 61 in Python 3.8.8 62.
The primary analyses consisted of two steps, with the first step relying on machine learning and the second step using conventional statistics.The Python syntax for both the machine learning and conventional statistical analyses can be found in the Supplementary Materials.www.nature.com/scientificreports/ Step 1: machine learning To address the first research objective of identifying the most important predictors of students' subjective wellbeing, we used a machine learning algorithm to model the three elements of subjective well-being.The scikit-learn package was used to perform five tree-based ensemble machine learning algorithms to identify the top predictors of well-being.We used different algorithms including gradient boosted decision tree (GBDT), adaptive boosting (AdaBoost), ExtraTrees (ET), RandomForest (RF), and light gradient-boost machine (LightGBM).We compared the predictive accuracy of these five algorithms and selected the best among them.Mean Square Error (MSE) was used to determine the prediction accuracy of the model.Mean Absolute Error (MAE) was used to evaluate the differences between the prediction and true value.Lower MSE and MAE values indicate a higher rate of model accuracy.The coefficient of determination (R 2 ) explains the amount of variance in well-being accounted for by the predictors.Among the different machine learning algorithms, LightGBM performed better than others with the lowest MSE and MAE values and the highest R 2 (see Table S1 in the supplementary file for more details).Therefore, we used the LightGBM algorithm as the primary analytic method in the first step.A tenfold cross-validation with 10 repeats was performed to streamline the models and select the top factors that have the strongest power for predicting well-being.For a better interpretation of the LightGBM model, we used the Shapley Additive exPlanations (SHAP) values that evaluate the contribution of each factor, not just the quality of the prediction itself 63 .
Step 2: conventional statistics To address the second objective of exploring how much variance in well-being was explained by the different predictors, we used conventional statistics.More specifically, hierarchical linear modeling (HLM) was conducted due to the nested nature of the data as the students were nested within schools, which were nested within countries/regions 64 .Life satisfaction, positive affect, and negative affect were the outcome variables.
The top predictors identified by LightGBM were designated as the predictor variables.Hence, rather than using all 37 predictors, we only used those predictors that emerged as important in Step 1.We calculated the fixed and random effects of all top factors at level 1. Random effects of schools and countries/regions were at level 2 and level 3, respectively.The value of the intraclass correlation coefficient (ICC) was used to examine the percentage of the variance in subjective well-being explained by the school and/or country level.The equations for the HLM models can be found in the Supplementary Materials (see Eq. S1).

Supplementary analysis
Supplementary analyses were also conducted to determine whether the results were similar or different across cultures.We analyzed the results separately for each of the eight cultural contexts.

Preliminary analyses
The descriptive statistics, variable description, and correlations with well-being for the overall sample can be seen in Table 2.The bivariate correlations among all variables are shown in Table S3 in the supplementary file.

Step 1: machine learning
The LightGBM regression model with 37 predictors was used as it performed better than the other machine learning algorithms such as GBDT, AdaBoost, ET, and RF.The comparison among the different machine learning algorithms can be found in Table S1.
The LightGBM regression models yield MSE values of 4.248, 0.195, 0.254, and can explain 33.6%, 37.3%, and 29.5% of the variance in life satisfaction, positive affect, and negative affect, respectively.Ten-fold cross-validation was performed.The step-by-step performance of models with an incremental number of factors are shown in Fig. 2. The models with the top 5 predictors had the lowest prediction error (i.e., MSE).This was true for all three dimensions of well-being.The optimal models with the top five factors explained 31.2%,35.3%, and 26.9% of the variance in life satisfaction, positive affect, and negative affect with MSE values of 4.404, 0.201, 0.263, respectively.
3 shows the ranking of the factors using the SHAP importance plot.Life satisfaction was best predicted by meaning in life, school belonging, parental support, fear of failure, and GDP per capita.Positive affect was Table 2. Descriptive statistics, variable description, and bivariate correlations with subjective well-being.**p < 0.001, r pertains to the correlation coefficient between the predictor and the outcome variables.All the labels, except for GINI and GDP per capita, were derived from the PISA Assessment and Analytical Framework.www.nature.com/scientificreports/most strongly predicted by resilience, meaning in life, belonging, parental support, and GDP per capita.Negative affect was best predicted by fear of failure, gender, experiences of bullying, school belonging, and meaning in life.Table 3 shows top predictors for each culture.

Label
Step 2: conventional statistics Table 4 shows the parameter estimates and p-values calculated from the HLM analyses.The value of ICC ranged from 0.02 to 0.04 and 0.02 to 0.07 for the school level and the country level, respectively.

Supplementary analyses
To explore whether the results across cultures were similar or different, we repeated the LightGBM regression analysis for each of the eight cultural groups.In general, the results in each of the eight cultural groups were www.nature.com/scientificreports/broadly consistent with the overall results.More detailed results can be found in the Supplementary Materials (see Table S2).

Discussion
In this study, we aimed to identify the most important factors predicting students' subjective well-being globally and across different cultural groups.Rooted in the bioecological theory, our model identified the top predictors of life satisfaction, positive affect, and negative affect.Life satisfaction was best predicted by meaning in life, school belonging, parental support, fear of failure, and GDP per capita.Positive affect was most strongly predicted by resilience, meaning in life, belonging, parental support, and GDP per capita.Negative affect was most strongly predicted by fear of failure, gender, experiences of being bullied, belonging, and meaning in life.Among the different predictors, school belonging and sense of meaning emerged as the most consistent predictor of the different dimensions of subjective well-being.

Contextual factors
Regarding context, parental support emerged as a crucial factor in predicting subjective well-being.Support from parents can facilitate students' positive self-evaluations and help them adjust to the school environment effectively 69 .
In terms of the school factors, our results suggest that the sense of belonging in school and experiences of being bullied were particularly important for subjective well-being.These findings also corroborate prior studies 43 .The need to belong is a basic human need 42,70 .Students who feel respected and safe in school tend to engage in school activities with more positive emotions, school satisfaction, and experience less negative emotions 42,71 .
Regarding the experiences of being bullied, our study found a positive association between bullying and negative affect.This finding is consistent with previous studies, which suggested that bullying is a critical negative experience that undermines students' well-being 17,72,73 .This is an area of concern as bullying might be especially acute in secondary schools 40,74 .

Implications
This study has several important theoretical and methodological implications.In terms of theory, the current study harnessed the power of a large-scale dataset that involved students from across 71 regions across eight cultural contexts.It provides a comprehensive understanding of the myriad predictors of students' subjective well-being across the globe.It also extends prior research which has mostly drawn on data from Western cultures.Furthermore, it helps highlight which among the diverse range of factors are most pertinent to predicting and explaining students' well-being.Although prior studies might have identified certain factors associated with students' well-being, the novelty of our study was the integrative approach we used.We compared a relatively large number of variables and identified the most powerful and salient predictors.
Methodologically, this study demonstrated the potential utility of combining both machine learning and conventional statistics in data analyses.Our findings suggested different key factors as most important for predicting different dimensions of subjective well-being, indicating the need to simultaneously consider different elements of well-being.Furthermore, it is important to note that not all predictors of well-being are created equal, some have better predictive power than others.However, comparing different well-being predictors in a single study is still relatively uncommon, as most researchers typically focus on the variable they are interested in, neglecting other variables that are also theoretically related to the outcome.
This study also has practical implications and pinpoints several variables that could become intervention targets.We focus on implications for school belonging and meaning, which emerged as consistent predictors of the different well-being dimensions.
Evidence-based interventions can be implemented to promote students' school belonging.For example, programs that reduce bullying in schools and those that foster cooperative learning and peer tutoring seem to be effective at enhancing school belonging 75 .Furthermore, when teachers show care for their students and create inclusive climates for their classes, school belonging is also enhanced 19 .Specific practices to support belonging could include providing opportunities for student participation, offering constructive feedback, and building positive teacher-student relationships.
Sense of meaning also emerged as a top predictor.Students who see themselves as part of something larger than themselves have a better sense of meaning.Meaning can be fostered through doing volunteer work, being part of extra-curricular activities, and participating in community service.A sense of meaning can also be enhanced when teachers try to help students see the relevance of what they are learning to their personal lives 76,77 .

Limitations and directions for future research
Despite its strengths, this study also has some key limitations.The first limitation is the cross-sectional nature of the PISA dataset.Hence, we are unable to explore the temporal and causal relationships among the variables.For example, is it the case that a higher level of school belonging at Time 1 leads to higher well-being at Time 2 or is the reverse also true? Longitudinal and experimental studies are needed to resolve these questions of directionality and causality.Second, we only focused on subjective well-being in this study.However, there are dimensions of well-being such as financial, social, and physical well-being 78 .Future studies can also include these other dimensions of well-being.
Third, this study focused on identifying the key predictors of subjective well-being but did not shed light on how these factors relate to or interact with each other.Future studies that explore mediation and moderation mechanisms might be needed to understand the nature of the relationships among the variables.
Last, it should be noted that PISA focuses on adolescent students.Therefore, the findings might only be limited to this developmental stage.Studies that cover other age groups are needed for a fuller account of wellbeing across developmental stages.

Conclusion
The present study examined the most important factors that predicted students' well-being.Across the 37 variables, school belonging and sense of meaning emerged as the most consistent predictors for all three dimensions of subjective well-being.The findings are generalizable across cultural contexts.Perhaps policymakers and educators can take cues from this study to identify potential intervention targets in their attempts to enhance students' well-being.

Figure 1 .
Figure 1.Conceptual model for the current study.

Figure 3 .
Figure 3.The top predictors of subjective well-being.

Table 1 .
Countries/regions within each culture.B-S-J-Z refers to Beijing-Shanghai-Jiangsu-Zhejiang which are all part of Mainland China.SAR refers to Special Autonomous Region.

Table 4 .
Hierarchical linear models predicting subjective well-being.a− 0.0000067, b − 0.00000083, c 0.0000014; Gender: Female = 1, Male = 2; ***p < 0.001.For instance, teachers can state how curricular content can be applied to daily life.They might also encourage students to make explicit linkages between what they are learning in class to their daily lives.